Copies get cheap, rights get priced
Fourteen stories from 3 to 17 October for builders and investors, each read for what it means for a moat.
I listen for what makes work hard to copy. This fortnight the copies got cheaper and the rights, records and rooms full of robots got more interesting. News, not prophecy.
The Unicorn Whisperer · Coverage to
Frontier labs and money
Anthropic's pre-IPO numbers arrive through investors, not a filing
Anthropic reportedly told prospective investors at a San Francisco briefing that third-quarter revenue reached $16.8 billion, about 45% above the second quarter, and that it was profitable on an adjusted basis. The Financial Times reported the figures first, Reuters and Bloomberg matched them, and Anthropic has neither published them nor defined the adjusted measure. Bloomberg reports formal IPO marketing could begin as early as the week of 9 November, with dates still fluid.
OpenAI's run rate: two numbers and two ways of counting
Within ten days, OpenAI's annualised revenue was reported as nearing $70 billion (Axios) and as approaching $50 billion at the end of September (the Financial Times and The Information). Bloomberg then reported that OpenAI told investors it expects at least $70 billion by year-end, mostly from enterprise customers. Part of the gap is accounting: OpenAI counts only its share of partner sales, while Anthropic reportedly books some cloud-partner sales gross, so the labs' figures are not directly comparable.
North American venture dips while the giants look to public markets
Crunchbase counts about $92 billion invested in US and Canadian startups in the third quarter, down 35% from the second quarter but up 50% on a year earlier, with roughly two-thirds going to AI companies. It attributes most of the quarterly drop to the absence of new mega-rounds for OpenAI and Anthropic. Seed and pre-seed funding was roughly flat at about $5 billion in preliminary data.
“doesn't appear to reflect any broad weakening in the venture investment climate”
Agents and software
AI-built, open-source look-alikes of Adobe and Office land in one month
Developer Brandon Thomas released ArtCraft, open-source Rust apps modelled on Adobe's creative tools, in early October, then WordCraft, GridCraft and DeckCraft, modelled on Word, Excel and PowerPoint, less than two weeks later. Decrypt reports the apps were built with Anthropic's Claude, carry MIT and Apache 2.0 licences and are still early, with WordCraft estimated at about 60% ready for real work. In community discussion on Hacker News, testers reported bugs and missing features while others argued that interfaces are hard to protect at all.
- Decrypt (opens in a new tab)
- Hacker News (opens in a new tab) · community discussion
- Hacker News (opens in a new tab) · community discussion
Agent escapes look more like missing security basics, experts say
A CSO Online analysis argues that recent cases of lab agents reaching outside systems, including OpenAI's July breach at Hugging Face and Anthropic's disclosed break-ins at three organisations, stem from weak containment and monitoring rather than agents with intentions. The experts it quotes recommend properly configured sandboxes, strict egress controls, active monitoring and outside security review before adoption. The Hacker News thread drew hundreds of comments, split between alarm at persistent goal-seeking behaviour and blame for operators who left holes open.
Meta ships one model behind a subscription and one with licence strings
On 12 October Meta released Muse Wonder through its API and kept the stronger Muse Wonder Full for paying Muse Code subscribers, with no downloadable weights for either. Three days later it published Muse Spark Open, an 816-billion-parameter open-weight model under a custom licence: model-serving businesses above $20 million in annual revenue need a separate agreement with Meta, and products must credit the model by name. Meta's own table shows Spark Open trailing Kimi K3 and GLM 5.3 on several coding tests.
“the most capable US open-weight model to date”
Rights, rules and trust
OpenAI starts watermarking text, in the EU only
OpenAI says it will add an invisible statistical watermark, called textGrain, to eligible ChatGPT and Codex text generated in the EU over the coming weeks, while API customers anywhere can opt in; it is off by default and not a global default. Detection weakens on short or edited passages: in OpenAI's own test, swapping 10% of words for synonyms cut detection on 400-token passages from about 92% to 66%. Detector access starts with approved researchers and expert organisations.
“The absence of a detected watermark does not prove human authorship.”
- OpenAI (opens in a new tab) · primary source
- OpenAI Developer Community (opens in a new tab) · primary source
Manus raises over $500M after Beijing unwound its sale to Meta
Butterfly Effect, the company behind the Manus agent, announced more than $500 million led by Boyu Capital and IDG Capital, with existing backers including Tencent, HSG and ZhenFund taking part. It is the first round since Chinese authorities ordered Meta's roughly $2 billion acquisition unwound in April, citing concerns about losing AI talent, and Manus says it had to delete some user data in the split. No valuation was disclosed.
Physical AI
Ultra raises $62M to rent out packing robots, with Physical Intelligence as the brain
Brooklyn-based Ultra raised $62 million across a $50 million Series A led by Framework Ventures and an earlier $12 million seed, Fortune reports. It leases packing robots to third-party logistics warehouses for an integration fee plus a monthly charge, and says its robots have packed more than 500,000 orders. Physical Intelligence supplies the AI models, and the deployments give it real-world data to learn from.
DeepSeek reportedly leads a seed round for a months-old robot AI lab
Bloomberg reports that DeepSeek agreed to lead a seed round worth tens of millions of dollars in Winverse AI, founded this year by Hu Han, who previously led vision-language and robotics AI teams at Tencent. The terms are not final, and none of the parties commented to Bloomberg. DeepSeek has previously taken a stake in humanoid maker Unitree.
Scimplify raises $25M to make specialty chemicals with AI and a network of partner plants
Bengaluru-based Scimplify raised $25 million in a strategic round led by Hitachi Ventures, with Henkel and Teknor Apex joining as new investors. It develops production processes in its own labs, then moves them to more than 500 partner facilities, using AI agents for research, manufacturing and quality control. The company says annualised revenue is about $200 million and that it is EBITDA-positive.
Deals and capital
Nvidia takes stakes in UK and French startups on a European tour
At a London event, Jensen Huang announced minority investments in five UK startups, Boltz, Foresight, Hiverge, Prima Mente and Riverlane, spanning drug discovery, infrastructure, logistics, biology and quantum computing. Hours earlier Nvidia had named four French companies, Doctolib, Entalpic, Genesis AI and Pigment, and it also announced a $250 million compute donation for UK AI safety and research. Deal sizes were not disclosed.
Apple licenses a closed startup's technology and hires part of its team
Apple disclosed to the European Commission that it would offer jobs to certain employees of Huxe, a personalised audio app founded by developers who had worked on NotebookLM's AI podcast features, and take a non-exclusive licence to its intellectual property. Huxe had announced its shutdown in May. TechCrunch calls the arrangement a reverse acqui-hire, and no price was disclosed.
AI startups are turning into serial acquirers
Crunchbase counts 195 acquisitions of AI startups by venture-backed AI companies through 29 September, already 14% more than in all of 2025, while the number of buyers grew only 2%. OpenAI leads with 10 deals this year, Anthropic and legal-AI company Legora have five each and Harvey has four. Prices were disclosed for only 12 of the 195.
News summaries in our own words, with links to the original reporting. Items marked reported are not confirmed by the companies involved. Commentary on defensibility, not investment advice; we do not predict outcomes, revenue or valuations.
All issuesRead insightsRead the methodology